The benefits of the Velvet Revolution in Armenia: Estimation of the short-term economic gains using deep neural networks

In March 2018, Armenia entered a new era of democracy after a non-violent revolution occurred on the streets of Yerevan. Locals called the revolution ‘velvet’, as, during the governmental transition, there were neither deaths nor violence. According to many prominent scholars, democracy comes with many benefits for the economy in the long run; however, in the first period immediately after the revolution, the economy also faced many challenges and risks. However, it may also be true that any revolution and rapid political change may lead to a short-term economic downturn, due to uncertainty, devastation or the destruction of public goods as well as a tendency on the part of investors to avoid conflict and areas of risk.

In this article, both the short-term costs and gains of the non-violent revolution on the economic situation in Armenia are analysed. The main hypothesis is that in the first year after the revolution in Armenia, the economic benefits outweigh the costs.

Three key elements have changed positively in Armenia since the revolution:

the level of democracy has increased;

All the three factors will have many positive effects on economic growth in the long term. Yet, conversely, two elements resulted in short-term costs for the economy:

the damage of public goods and labour inactivity during the revolutionary process and

In addition, according to some scholars, democracy may lead to an increase in consumption, thus decreasing savings and investment levels, also resulting in a negative impact in the long term.

This article primarily aims to analyse the short-term economic impact of the non-violent revolution in Armenia. Since this revolution raised the level of direct democracy in Armenia, this study shows the short-term economic impact of democracy. Even though there are many publications about the impact of democracy on the economy in general (for the long-term) or the impact of the revolution on the geopolitical situation in the country, there are hardly any articles discussing the short-term economic costs or benefits of the revolution or establishment of democracy. This article tries to fill this gap by estimating the short-term net economic effects of the revolution, taking the example of Armenia and using advanced methodologies such as machine learning (ML) and deep neural networks (DNNs).

In addition, this article compares the DNN and ensembling model prognosis with more standard, classic model approaches, such as autoregressive integrated moving average (ARIMA) or error correction model (ECM), using the example of the Armenian economy. The growing number of articles using the DNN models in macroeconomic estimations indicates the potential of these methods and creates a new opportunity for projections and forecasts in macroeconomics. Moreover, some articles indicate that the DNN models for developing economies are better at capturing non-linear links between the potential influences of external factors and GDP. This study measures the accuracy of forecasting for all the aforementioned models and uses the most accurate models to measure the net cost/benefit of the revolution, based on the 95% confidence interval of the prognosis.

In this article, the quarterly data of GDP and 18 explanatory variables from the official webpage of the Statistical Committee of the Republic of Armenia (SCRA) and International Monetary Fund (IMF) are used for the years 2000–2018. Data are transformed into time series (TS) and seasonal autoregressive integrated moving average (SARIMA), ECM, recurrent neural networks (R NN), convolutional neural networks (CNN) as well as ensembling models [simple average, least absolute shrinkage and selection operator (LASSO) and linear stacking] are implemented to forecast the period 2018-Q2–2019-Q1. Later, it is compared with real data, and the difference will indicate either a net surplus or the cost of the revolution.

Section 1 is dedicated to a short, historical background of Armenia and its political and economic situations. It also discusses the main drivers of the Armenian economy and its dependency on its strategic partner, Russia. Section 2 discusses the political and economic events that led to the national revolution, highlighting the motives behind the movements and the main benefits, based on public opinion surveys and different international ratings. The attitude of the people towards the new government and its tendencies is also considered as well as the key changes made by the new government immediately after the revolution. Finally, in this section, the impact of democracy on economic performance is discussed by analysing various studies regarding this topic.

Taking Armenia as an example, Section 2 is devoted to the analysis of short-term economic growth after the revolution and its comparison with predictions of economic activity during the same period, using pre-revolutionary data. Economic activity has been measured by GDP, which was predicted using the historical data of GDP, commodity prices, interest rates, inflation rates, foreign direct investments (FDI), unemployment, debt to GDP ratio and other factors, discussed in Section 2. Both classical (SARIMA, ECM) and non-linear DNN models (RNN, CNN) as well as the ensembling of all models were used to analyse 53 quarters of data (2000-Q1–2013-Q1), then data for the next 4-year period (2013-Q2–2018-Q1) were analysed. Afterwards, using those models, the GDP of Armenia was forecasted quarterly for the period 2018-Q2–2019-Q1. Finally, the prediction of 95% confidence intervals was estimated and compared to actual, post-revolutionary values. Since predictions are generated by using pre-revolutionary data relating to internal factors, which are a consequence of the previous regime and external factors (commodity prices, which are independent of Armenian political change), prediction estimates will represent what would have been the economic growth had there been no revolution, and the difference will indicate the approximate effect of the revolution on the economy.

As no papers are analysing the economic situation in Armenia after the revolution and few articles are analysing the economic impacts of the political revolution, this article tries to bridge the gap, providing econometric evidence using neural networks and ensembling models.

As one of the oldest civilisations in the world, Armenian history dates back to the twelfth century BC. Since then, the Armenian kingdoms and dynasties have been known for their wealth and power; however, after the invasion of the Arabic, Mongolic and Turkish tribes in the thirteenth century, Armenia lost its sovereignty until the twentieth century. In 1918, Armenia gained its independence but became part of the Soviet Union 3 years later as an Armenian Soviet Socialist Republic (Yeghiazaryan, 2014, p. 42).

The sectoral structure of the current Armenian economy has been modelled on the structure of the economy of the Armenian Soviet Socialist Republic. Since the economy of landlocked Soviet Armenia was relying heavily on the network of other member states of the Soviet Union, especially in the industrial sector after the collapse of the USSR and the integration of the capitalistic system, not all sectors of the economy continued to exist and the economy went into recession.

Due to the collapse of the USSR, Armenia overcame an economic crisis in the energy sector. During the years 1990–1993, the GDP of Armenia almost decreased twice with GDP in 1993 is only 46.9% of that of 1990. Power generation also dropped by 46%, due to the blockade of transportation, decrease in fuel imports as well as the closure of nuclear power plants. During these years, most of the regions did not have access to electricity and natural gas (Tagharyan, 2018, p. 96). In 1994, the recession phase was replaced by recovery and except for the year 2009, GDP has experienced positive growth ever since; however, the structural quality of the economy has not improved. During the 1990s, the emigration rate from Armenia reached its peak due to the difficult economic situation and ever since, more people have left the country than have arrived (CIA, 2019). However, the situation has changed since the revolution, as during 2018-Q2–2019-Q1, more Armenians returned home than left (Pashinyan, 2019).

2.1

Main drivers of the Armenian economy

Some of the main factors of growth during the late 1990s and early 2000s were international grants, loans and investments in strategic sectors of the economy (energetics, transportation and telecommunications) and the increasing volume of private transfers. The latter mostly came from the growing Armenian diaspora (EV RC, 2005). During this period, the share of agriculture in the GDP grew, while the share of the industry dropped significantly (Yeghiazaryan, 2014, p. 49).

Figure 1 represents the quarterly Armenian GDP for the period 2000–2019. The data from 2003 to 2008 indicate that the economy moved into an expansion phase, as during that time, the GDP growth rate reached two digits. However, as the economy depended heavily on investment and transfers from outside the country and had an imbalance within its sectors, the global financial crisis had a very negative impact in 2009, when Armenia experienced a 14.1% decrease in economic activity, and since then, recovery has been very slow. In recent years, two of the main contributors to economic growth have been agriculture and the metallurgical industry. The growth of the metallurgical industry could be explained by increasing exports and the growing prices of commodities. Even now, the Armenian economy is heavily dependent on FDI and private transfers from abroad, which, in the long run, make the economy volatile.

Fig. 1

2.1.1

Agriculture

The main attributes of agriculture in Armenia are low productivity and a very high dependency on natural factors. During the years 1993–1994, 90% of agricultural land had been privatised, resulting in the number of people engaged in that sector doubling, yet productivity during the recession decreased. With limited access to basic resources, such as electricity and natural gas, people were generally using their land as a means of survival and primarily for personal use. However, after the recession, the situation did not change a great deal, as agricultural land was still divided into very small pieces with 96.7% of the land belonging to small private farms and only 3.3% to middle and big agricultural trade organisations. According to data from 2012, the average size of farms was 13,700 m², whereas the total area of arable land was 20,500 km². Eighty-eight per cent of the farms were smaller than 20,000 m² and they made up 77% of the total arable land (Yeghiazaryan, 2014, p. 47). Another problem was the lack of any insurance system to protect against natural disasters and the influences of climate, which was only partially subsidised by the treasury of the government. This has led to more instability in the agricultural sector and the total economy as a whole (as it was one of the main contributors to economic growth). For instance, during the financial crisis, agricultural activity only dropped by 0.1%; yet, in the following year, it dropped by 16%, primarily due to the severity of the climate.

2.1.2

Industry

The mining and metallurgical industries are among the fastest growing industries and are the main contributors to economic growth. In Figure 2, the first pie chart represents the structure of the industry in Armenia and the second pie chart illustrates the composition of process manufacturing. In 2017, according to the Statistical Committee of Armenia, more than a quarter of the industrial volume was derived from those two sectors, both of which have a large share of exports.

Fig. 2

Another important sector is the food industry, which comprises 18% of the total industrial volume. However, most food produced are consumed in the local market, and this sector heavily depends on agriculture.

2.1.3

Energy sector

The energy sector is mainly dependent on Russian companies and investors. Most of the oil imports come from Russia. The supply of natural gas is controlled by ArmGazProm, an Armenian–Russian joint venture, 20% of which was owned by the Armenian government and 80% by Gazprom (Russia) until 2013, after which time Gazprom took ownership of 100% (Arzumanyan & Abovyan, 2014, p. 4). Even though a third of natural gas imports come from Iran, the operation is fully controlled by a company majority owned by the Russian government (Gazprom, 2018). In addition, the Metsamor nuclear power plant is fuelled and has been controlled by the Russian company Rosenergoatom since 2003. Full control was given by the government to pay off debts raised by nuclear fuel imports (WNA, 2019). Therefore, the whole sector is controlled by Russia, making the Armenian economy heavily dependent on Russia.

As indicated in the introduction, the main hypothesis of this article is that the positive impacts of the revolution in Armenia can even be seen within the first year. To validate this hypothesis, the article uses an econometric model, which measures the cost (benefits) of revolution in terms of GDP for the first year after the revolution (Q2-2018–Q1-2019) by forecasting this period using only pre-revolutionary data (until Q1 2018), simulating the scenario as if there were no revolution. Afterwards, the outcome is compared to real, post-revolutionary data. The difference between the simulation and real data indicates the net impact of the revolution on the Armenian economy for the first four quarters. The effect of the revolution is measured as a difference in the changes in the sum of GDP after the revolution and the corresponding sum for the counterfactual state.

To avoid overparameterisation, the pre-revolutionary dataset is divided between training and testing datasets: the model is trained based on 53 observations (Q1 2000–Q1 2013), later validated according to 19 pre-revolutionary observations (Q2-2013–Q1-2018). All assessments of validity and quality of the prediction methods are based on the validated set of observations.

To calculate the prognosis of GDP, this article uses seven models: two DNN models (long short-term memory recurrent NN and convolutional NN), ECM, SARIMA and three ensembling models (simple average, stacking linear and LASSO).

3.2

DNN models: CNN and long short-term memory

In recent years, some scholars have been using unsupervised learning models, like DNNs, to predict or forecast GDP. Tkacz (2001) found that while using quarterly basis data, NN models do not outperform linear models, yet in the long term (using yearly basis data) unsupervised models are superior to linear models. He showed that the non-linear influence of monetary variables seems to be more relevant over a longer period. Qi (2001) and Jahn (2018) confirm this view for developed countries. Chuku, Odour, and Simpasa (2017), using the example of frontier economies in Africa, show that NN models also perform better than the ARIMA and other structural econometric models for developing countries. The authors draw three main conclusions:

Neural network models can be used to predict GDP growth in countries with developing economies ‘that are exposed to potential chaotic influences from commodity prices, external factors and even political economy factors because this class of model is better able to learn the system and capture the nonlinearities inherent in the input variables’.

It is recommended to revalidate NN predictions with structural economic models, as in some cases it can produce outliers in certain data points.

In their paper, structural econometric models underperformed in the case of developing African countries mainly ‘because of the sudden changes and chaotic patterns of macroeconomic variables in developing economies’.

Cook and Hall (2017) forecasted the US unemployment rate using four different types of NN methods: FC network, CNN, RNN and encoder-decoder network. In the last two models, they use long short-term memory (LSTM) as an architecture. The authors exhaustively explain the theory behind all these models and as a comparison, use the VAR model. As a result, all deep learning models perform better than benchmark models in the short term (one or two-quarters prognosis), yet, for the longer term (three or four quarters prognosis), only encoder-decoder network outperforms the others. Using the same US unemployment data, Stasinakis et al. (2015) carried out another comparison of NN models with linear models. They discovered that the radial basis function neural network performed statistically better than the rest of their models. However, in both of these articles, a univariate TS was investigated, while in contrast to some linear TS models, such as VAR or ARIMA, NN models can also ‘learn’ from numerous features using multivariate TS data.

There are various types of Neural Networks, however, only a few of them suited the multivariate TS models. More advanced networks are called DNNs. Compared to simple shallow NN models, DNN has more than two layers and uses specific mathematical modelling in each layer to process data. Both CNN and RNN (LSTM) that have been used in this study are part of that ‘family’. Both models have been created in Python by using the ‘keras’ library. All parameters were selected to minimise the mean absolute error (MAE) of the outcome.

3.2.1

Convolutional neural network

CNNs have received a lot of attention in the ML community over the last few years, due to various successful applications in areas such as object detection, speech recognition and TS prediction. The original goal of CNN was to form the best possible representation of visual images to recognise them. The CNN solution needs to detect and classify and be sufficiently robust. Unlike the fully connected setting in other NN models, in the convolutional layer of this network, each input is connected in the layer to a few computational nodes. Each node receives input from a smaller clustered set of input nodes. Later, all weights of input in the layer are shared across all computational nodes, identifying patterns that increase predictive accuracy (Cook & Hall, 2011).

Figure 3 represents the architecture of CNN implemented in this article. An input model takes a 12x4 matrix with 12 features (also including the GDP growth rate) for the periods t-3, t-2, t-1 and t and returns an output of GDP growth rate for period t+1. As the dataset is not very big (having only 76 quarters) including more time steps would have reduced the training set. To get the best results, two convolutional layers are stacked together. To reduce the layers of each stack, max pool filter (sample-based discretisation process, reducing the dimensionality) is applied, and then it is flattened into a one-dimensional layer for the fully connected layers. The output of the final phase gives the prediction of the target variable for the t+1 period.

Fig. 3

3.2.2

Recurrent neural network (LSTM)

The traditional neural network takes each input as an individual data point, assuming the data are nonsequential and each data point is independent of the others. In contrast, RNN is designed with loops that take not only one new input at a time but also the output from a previous data point that was fed into the network. Therefore, at a specific period of k, to predict GDP_k+1 as an input model, it not only takes x_k (including lagged values) but also the output of the previous state weight matrix W_k−1 between the input and the hidden unit (for period t − 1). As a result, RNN can remember the analysis that was carried out up to a given point, by maintaining a state as a memory. That state transfers back into the neural network with each new input to predict the output.

The LSTM model is a type of a recurrent NN model, which, in comparison with the vanilla RNN model, is capable of maintaining a strong gradient over many time steps. This means that the model can be trained with relatively long sequences, along with the short-term memory of the most recent network outputs. The architecture of the LSTM unit consists of the following elements—memory cell and three logistic gates: input, forget and output. The gates are logistic functions of weighted sums, where the weights might be learned during iterations. The input and forget gates manage the state of the cell or so-called long-term memory. The output gate produces the hidden state or the output vector. (Petneházi, 2018, pp. 2–3).

Figure 4 represents the architecture of the model used in this study. Once again, the model has a 12x4 matrix as an input (12 features for four previous periods). The first part of the model is an LSTM network, consisting of four units, (as input data have four-time steps) and 800 nodes. Twelve features from the t − 3 time step are passed to the first unit of the network, which uses the random initialised hidden state and output to produce the new hidden state and first step output. The next unit takes three inputs: the output and hidden layer from the previous unit and the features from the next time step. The second part of the model consists of fully connected layers with 50 nodes of the hidden layer and output layer.

Fig. 4

3.3

Validation models: SARIMA and ECM

To compare and validate the DNN models, two structural econometric models are included in this article. Both are widely used in literature for prediction purposes.

3.3.1

Error correction mechanism

First is the linear approach using error correction mechanism (ECM). To build the model, Engle-Granger's two-step method (Engle & Granger, 1987) has been implemented. At the start, the integration level of all variables is checked, and then the linear model with lagged variables is made to check co-integration between variables of the same integration level. The model has been transformed from general to specific, leaving only statistically significant variables with their time lags. Afterwards, the residuals from the OLS regression are calculated from the model, and the fourth lag of the last residual is included in the final model. The residual is taken from 1 year ago (fourth lag) so that it is possible to predict four periods using only pre-revolutionary data. The final model has the following formula: ΔGDPt=β1ΔGPDt−4+β2Inft−4+β3Inft−5+β4Got−2++β5Cot−2+β6Mot−1+β7ε⌢t−4+vt\matrix{{\Delta {\rm{GDP}}_{t} = {\beta _1}\Delta {\rm{GPD}}_{t - 4} + {\beta _2}{\rm{Inf}}_{t - 4} + {\beta _3}{\rm{Inf}}_{t - 5} + {\beta _4}{\rm{Go}}_{t - 2} +} \hfill \cr {+ {\beta _5}{\rm{Co}}_{t - 2} + {\beta _6}{\rm{Mo}}_{t - 1} + {\beta _7}{{\mathord{\buildrel{\lower3pt\hbox{$\scriptscriptstyle\frown$}}\over \varepsilon}}_{t - 4}} + {v_t}} \hfill \cr} where ΔGDP is the differentiated GDP, Inf is inflation, Go, Co and Mo are gold, copper and molybdenum commodity prices, respectively. If commodity prices have external parameters and the Armenian revolution has no impact on them, there is no need to restrict them to have fourth lag or higher to predict post-revolutionary quarters.

3.3.2

Seasonal autoregressive integrated moving average

Due to the seasonality of the data, another favourable method is capturing the trend line by using SARIMA. The SARIMA TS approach to forecasting is one of the most popular approaches, due to its ability to use past values without any requirement for economic fundamentals and theory (De Gooijer & Hyndman, 2006).

To create an appropriate ARIMA model, ACF and PACF graphs are built as a ground point and are represented in Figure 5. According to the graphs, the model has some AR and seasonal effects.

Figure 5

The final model has been built in R using ‘auto. arima’ function from the package ‘function’, which uses the algorithm of Hyndman and Khandakar, which minimises selected information criteria and MLE matches the best ARIMA model (Hyndman & Khandakar, 2008). According to the algorithm, the best structure was SARIMA(0,0,1)(0,1,1)₄. After fitting the model, the Ljung-Box test on residuals was checked. As the test p-value is greater than 0.05, the residuals are independent of the model. In addition, this model has the lowest MAE on the training dataset compared with other ARIMA models tested during this study.

To statistically evaluate the forecasts from all models, the root mean squared error (RMSE), mean absolute error (MAE), median absolute error (MedAE), standard deviation (SD) and R² of each testing set are compared. For the first three statistics, the smaller the outcome, the better the performance of the model and vice versa R². Apart from R², all values are shown in millions of AMD.

Table 1 represents the outcome of the statistics mentioned above arranged in terms of MAE results. According to the results, CNN outperforms the rest of the models, except for the stacking linear model that shows a slight difference. Those two models have similar prediction accuracy because, in the training dataset, CNN also outperforms the rest of the models, hence recording the biggest parameter for the stacking model. The LSTM and ECM models had worse prediction accuracies, which is why the simple average model also underperformed compared with the other ensembling LASSO and CNN models. According to this outcome, CNN is the superior model regarding most of the statistical criteria. The ensembling models did not significantly improve the accuracy of the individual performances; however, they were more accurate compared to LASSO and ECM and to a certain extent SARIMA.

By comparing both the training and test prognosis on both neural networks, results showed that they performed in quite a similar way on the training dataset, yet on the test set, the LSTM model overpredicted the lowest values (Figure 6). The bad performance of the LSTM model was mostly due to the data frame size, as it was very small in comparison with certain big data TS analyses, where it overperformed.

Fig. 6

Figures 7–9 are graphical representations of forecast on the test set as well as the post-revolutionary year (shaded in grey). To consider the variability of the predictions, the prediction interval is introduced on those figures. A 95% confidence interval of each model prediction is shaded in a different colour to see how accurate they are and how they differ from one other. The confidence interval of prediction is calculated by the following formula: ΔGDP^t±z*σ\Delta {\widehat {GDP}_t} \pm z*\sigma where ΔGDP^t\Delta {\widehat {GDP}_t} is the predicted value for time t, z is the critical value from Gaussian distribution (which is 1.96 for the 95% confidence interval) and s is the standard deviation of the model based on the test set results (Table 1).

Fig. 7

Fig. 8

Fig. 9

From the graphs alone, even without any calculations, the overperformance of CNN over ECM (Figure 7) and SARIMA over LSTM (Figure 8) can be seen in the test set, yet the ensembling models have very similar performance, and it is even hard to distinguish between them in Figure 9. However, in the grey area, since 2018, Q2, there has not been any obvious dominance or similar pattern between models. CNN, LSTM and linear stacking models have predicted the real GDP to be lower, yet for the other models, almost all observations are within the prediction confidence interval.

If converted to Euros (the exchange rate is taken from SCRA as of 31 December 2018), the sum of the difference between real and forecasted values, considering the confidence intervals of prediction, reveals the value of 856.7 million EUR, based on the CNN model difference. Similarly, for linear and LASSO stacking, it is slightly lower (around 440.2 and 523.5 million EUR, respectively), with the SARIMA and simple average models, it is very close to 0; yet for LSTM, it is −571.5 million EUR and for ECM, −191.4 million EUR. However, considering that both LSTM and ECM models underperformed in the test dataset and other models, like CNN, have more predictive accuracy, this result is more biased compared to other models.

The main conclusion of this article is that the Velvet Revolution in Armenia did not have any negative impact on the economy during the first year after political turnover. When the difference in the sum of GDP after the revolution and the same sum for the counterfactual state were measured, the best models (based on their accuracy) out of seven indicated that the difference between the GDP prognosis using only pre-revolutionary data and post-revolutionary GDP was either insignificantly small or positive, which means that the benefits of the revolution outweighed the costs. Moreover, according to the most accurate model, CNN, the Armenian GDP gained a surplus of around 850 million EUR within a year, following the revolution.

Section 1 highlights all the threats and opportunities of the Armenian economy. After the collapse of the Soviet Union, the Republic of Armenia overcame many challenges, including economic occupation, lack of some vital resources (gas and electricity) and electoral fraud, and as a result, the country was left with a government of oligarchs. However, the Armenian revolution in 2018 proved that the people did not give up and they changed the political situation in the country, establishing direct democracy. After coming to power, one of the main goals set by the government was to fight corruption and the shadow economy, and within a year, it had added 130 million USD to the budget, for this purpose. However, abrupt changes are often accompanied by certain negative effects, such as labour inactivity or the damage of public goods during the revolution. This section also conducted a review of the literature highlighting the impact of democracy and political change on economic growth.

In the second section of this article, an empirical analysis is carried out, using NN methods to ascertain whether there were any costs\benefits to the economy following the revolution, compared to the situation in the country had there been no revolution. This section included all the theoretical background of each model, a literature review of NN models in macroeconomics and their implementation.

The results in this section support the idea that for developing countries, some NN models may work better than traditional models; however, it is still recommended to perform structural econometric models to have a point of comparison, as in the case of LSTM which might underperform heavily. In addition, results show that the ensembling models provide a less volatile forecast using information from all models. However, in terms of MAE accuracy, CNN overperforms all three ensembling models. The performance of those models could have increased by excluding the biggest underperformers, ECM and LSTM.

According to model accuracies, CNN is the superior model for this study. Based on this model, the Armenian economy was shown to have gained an additional 850 million EUR from the revolution. The budget surplus that the government gained to fight the shadow economy (130 million USD) only partially explains this surplus. Another explanation may be linked to other factors mentioned in Section 2, such as eliminating monopolies, removing barriers to establishing new businesses, increasing tourism, ensuring the competitiveness of markets and most importantly a more optimistic attitude among the population concerning the future and increased trust in the government. Finally, we can conclude that democracy and political turnover only bring benefits, given that during the revolution, there was no violence and most importantly, no geopolitical game was played between the other superpowers.

As a result, the most accurate models either capture a positive outcome post-revolution or no changes at all; hence, there are no grounds for rejecting the main hypothesis of this article, that is, the benefits of the revolution could be seen even in the first year after the political change.